Both versions are aligned block by block, in reading order: title, the essentials, then paragraph by paragraph. Where the translation merged or split a paragraph, the matching cell stays empty — we never pair two passages by guesswork.
Hugging Face Hubに並ぶ公開モデルの数が、243万件から296万件へ増えた。2026年1〜7月の動きを追ったHugging Faceの分析で、中国のラボや企業が、モデルの大きさとコミュニティでの広がりの両方で存在感を強めていることが浮かび上がった。
The number of publicly released models listed on Hugging Face Hub grew from 2.43 million to 2.96 million. An analysis by Hugging Face tracking activity from January through July 2026 found that Chinese labs and companies are strengthening their presence both in model size and in reach across the community.
規模で先行したのは中国勢だ。Hugging Faceによれば、中国のラボが公開したモデルの最大サイズは、対象期間のほぼ毎月、米国のラボを上回った。中国勢の月ごとの上限は7540億〜2兆7800億パラメータ。一方、米国勢は7カ月中5カ月で1300億パラメータ未満だった。Moonshot AI、MiniMax、Xiaomi、Z.aiは、10億や数百億パラメータのモデルを順に増やす従来型の展開を避け、700億パラメータ未満のモデルをほとんど公開していない。
China’s players took the lead in scale. According to Hugging Face, the largest model released by a Chinese lab was bigger than the largest model from a U.S. lab in almost every month of the period examined. The monthly ceiling for Chinese players ranged from 754 billion to 2.78 trillion parameters. By contrast, U.S. players remained below 130 billion parameters in five of the seven months. Moonshot AI, MiniMax, Xiaomi and Z.ai have largely avoided the conventional strategy of sequentially expanding models from 1 billion to tens of billions of parameters, releasing few models below 70 billion parameters.
対照的なのがTencentとAlibabaの「Qwen」だ。Qwenは10億パラメータ未満から大規模モデルまで幅広くそろえ、開発者が共通の土台として使えるモデル群を狙う。Hugging Face Hub上のQwen派生モデルは15万1448件に達し、Meta全体の2.6倍、「Llama」関連だけと比べると4.7倍だった。2026年1〜7月には、1日180〜210件のペースで派生モデルが増えた。
Tencent and Alibaba’s “Qwen” present a contrast. Qwen offers a broad range, from models with fewer than 1 billion parameters to large models, aiming to provide a family of models that developers can use as a common foundation. Qwen derivatives on Hugging Face Hub reached 151,448, 2.6 times the total for Meta and 4.7 times the number associated with “Llama” alone. From January through July 2026, derivative models grew at a pace of 180 to 210 per day.
広がりを支えたのは、定期的な公開、サイズの選択肢、Apache 2.0という利用しやすいライセンスの組み合わせだ。中国発で2026年に公開された200億パラメータ超のモデル178件では、59％がApache 2.0、22％がMITだった。ただし、Kimi K3やQwen 3.8のように、非商用制限や収益分配条件を含む例も出ている。Hugging Faceは、こうしたモデル公開の収益源をライセンス料ではなく、API、クラウド、ハードウェア、プラットフォームにあると分析している。
The combination of regular releases, a range of size options and the accessible Apache 2.0 license supported this expansion. Of 178 models released in 2026 from China with more than 20 billion parameters, 59% used Apache 2.0 and 22% used MIT. However, examples such as Kimi K3 and Qwen 3.8 also include noncommercial restrictions or revenue-sharing requirements. Hugging Face analyzed the revenue sources for these model releases as coming not from licensing fees, but from APIs, cloud services, hardware and platforms.
具体的に変わるのは、開発者がQwenのような基盤モデルを改変し、自分の用途に合わせた派生モデルを作りやすくなる点だ。一方で、ダウンロード数や派生モデル数は品質や商用採用、市場シェアそのものではない。実際、累計ダウンロードの83％は10億パラメータ未満で、1000億超は1％にすぎない。大型モデルの公開競争は注目を集めるが、ダウンロードで中心となっているのは10億パラメータ未満の小型モデルである。
The practical change is that developers can more easily modify foundation models such as Qwen and create derivative models tailored to their own needs. At the same time, download counts and the number of derivative models are not in themselves measures of quality, commercial adoption or market share. In practice, 83% of cumulative downloads went to models with fewer than 1 billion parameters, while those with more than 100 billion accounted for just 1%. The competition to release large models attracts attention, but downloads are dominated by smaller models with fewer than 1 billion parameters.